Fluently Lying: Adversarial Robustness Can Be Substrate-Dependent
A research paper, now withdrawn, explored adversarial robustness in object detectors, specifically focusing on a phenomenon termed "Quality Corruption" (QC). The study observed that one model, EMS-YOLO, a spiking neural network, retained a high percentage of detections while its accuracy collapsed under adversarial attack. This behavior, termed QC, was found to be substrate-dependent, appearing only in one of four tested SNN architectures and proving resistant to standard defense mechanisms. AI
IMPACT Reveals that adversarial failure modes can be specific to AI model architecture, challenging existing defense assumptions.